Research On Space Infrared Targets Recognition Based On Convolutional Neural Networks

The space target recognition based on infrared detection plays a key role in the missile defense system,which is of vital importance to the national security.With the rapid development of artificial intelligence and the requirement for intelligent warfare,establishing an intelligent recognition system based on data driving is an urgent and challenging problem.The dessertation studies the important and difficult problems of recognition systems on the basis of Convolutional Neural Networks(CNNs).The main work and contributions of the dessertation are the following four aspects:Firstly,the infrared radiation intensity sequence model of space targets is studied.By analyzing the temperature model,the orbital motion and the projection area model of space targets,the infrared radiation intensity sequence model is established.Based on the target micro-motion model,the relationship between the inertia parameters and the micromotion period is built.And the influence of the target inertia parameters on the attitude motion mode and infrared radiation intensity sequence is investigated.The main factors inducing the variation of the infrared radiation intensity sequence of space target are further discussed.Finally,the radiation intensity sequence of the target is simulated providing a basis for the succeeding researches.Secondly,a multi-scale convolutional neural network(MCNN)is proposed for the classification problem of long-distance detection of space targets with severe noise.The network incorporates multi-time scale and multi-frequency scale transformation.By means of the local convolution and full convolution,the low-level features and high-level features are utilized simultaneously to capture the long-term trend and local fluctuations of sequence,which explores more abundant features for classification.The performance is evaluated on the public UCR time series data and the target simulation data.The results show that the MCNN effectively improves the classification performance of the space target under severe noise pollution.Thirdly,a classifier based on the reconstruction of two-dimensional representation of sequences is proposed to solve the classification problem with data missing.Firstly,a two-dimensional representation model of infrared radiation intensity sequence is constructed.Then,the two-dimensional representation of the sequence with missing data is reconstructed by the denoising encoder.Finally,the MCNN is used for classification based on the reconstructed recurrence plots representation.The representation based on the recurrence plots establishes the temporal correlation in time series through the recurrence matrix,and explicitly characterizes the structural pattern of the target radiation intensity sequence and the dynamic characteristics of the system generating the target radiation intensity sequence.It has more stable performance compared to the original time series in sequence reconstruction and simplifies the feature learning of subsequent classification algorithms.The results on the UCR time series data and the target simulation data show that the recurrence plots representation is better in sequence reconstruction and classification.The sequence reconstruction based on the denoising autoencoder further improves the stability of the space target recognition system through missing value recovery.Finally,the interpretability and transfer learning of CNNs on the classification of target radiation intensity sequence is studied.To solve the problem of interpretability of CNNs on the space target recognition,a class activatation mapping technology is adopted to locate the regions in the original sequence that contribute to the classification,which helps us to better understand the decision-making process of the model.To solve the problem of adaptability of recognition system under a small number of training samples,a classification framework based on the attention mechanism model with transfer learning(Attention-CNN)is proposed.The model uses the attention mechanism to capture important information to construct a general feature extraction module.The transfer learning under similar data sets alleviates the feature learning under a limited training sample.Experiments were carried out on the UCR time series and the target simulation dataset.The results show that the classification performance of Attention-CNN is improved compared to the CNN without attention module.Under the limited training sample,the Attention-CNN with transfer learning on similar tasks achieve better classification performance with fine-tuning the last layer of network.